When demand paging is enabled, only a minimal set of functions
are available at boot. Anything not marked as boot functions
would not be loading in memory at boot which would result in
page faults when jumping to those functions. However, at early
boot, demand paging has not been enabled yet. So we need to
mark necessary functions as boot functions so they are placed
in the correct linker section where they are loaded at boot.
Fixes#56414
Signed-off-by: Daniel Leung <daniel.leung@intel.com>
Add the `zephyr,pm-device-runtime-auto` flag to `pm.yaml` and
`struct pm_device`.
This flag is intended to signify to the boot system that device runtime
PM should be automatically enabled on the device after the init function
has run.
Only run `pm_device_runtime_auto_enable` function on a device if
initialisation succeeded. This prevents actions being run on devices
that are not ready.
Signed-off-by: Jordan Yates <jordan.yates@data61.csiro.au>
Returning an error code when PM is not supported for a device only makes
writing drivers harder, as instead of checking for failures with `< 0`,
they also need to check for `-ENOTSUP`.
From the point of view of a driver, an `-ENOTSUP` return value is
equivalent to success, as if the device it is trying to turn on doesn't
support PM, by definition it is already enabled. This is equivalent to
the API behaviour when `CONFIG_PM_DEVICE_RUNTIME=n`.
Whether a device supports PM or not can still be determined at runtime
by inspecting the return code of `pm_device_runtime_enable` if
necessary.
Signed-off-by: Jordan Yates <jordan.yates@data61.csiro.au>
Only run `pm_device_runtime_put` on the dependent domain if the original
claim operation succeeded. This fixes unbalanced operations when running
```
pm_device_runtime_get(dev);
pm_device_runtime_put(dev);
```
On a device that does not have PM enabled but does have a power domain.
Signed-off-by: Jordan Yates <jordan.yates@data61.csiro.au>
Check if device runtime is enabled in the beginning of
runtime_suspend to avoid unnecessary checks.
This does not change the current behavior.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
Check if device runtime is enabled in the beginning of
pm_device_runtime_get to avoid unnecessary checks.
This does not change the current behavior.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
To be able to call pm_device_runtime_put() from an IRQ context, move
from mutexes to semaphores and force the async path when the put
operation is called from an ISR.
Signed-off-by: Carlo Caione <ccaione@baylibre.com>
For the async operation move from condition variables to events to
reduce the dependency on the mutexes that cannot be used in IRQ
context.
Signed-off-by: Carlo Caione <ccaione@baylibre.com>
Check that the device we are trying to use has successfully powered up
when it is on a power domain. If it has not, release the request for the
power domain to be powered to ensure we do not have dangling requests.
As a result of this, `PM_DEVICE_ACTION_RESUME` will not be run on a
device if `PM_DEVICE_ACTION_TURN_ON` has already failed.
Signed-off-by: Jordan Yates <jordan.yates@data61.csiro.au>
Return `-ENOTSUP` on calls to device_runtime functions if the underlying
device does not support power management.
Fixes#45648.
Signed-off-by: Jordan Yates <jordan.yates@data61.csiro.au>
In order to bring consistency in-tree, migrate all subsystems code to
the new prefix <zephyr/...>. Note that the conversion has been scripted,
refer to zephyrproject-rtos#45388 for more details.
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
A common pattern here was to take the work item as the subfield of a
containing object. But the contained field is not a k_work, it's a
k_work_delayable.
Things were working only because the work field was first, so the
pointers had the same value. Do things right and fix things to
produce correct code if/when that field ever moves within delayable.
Signed-off-by: Yong Cong Sin <yongcong.sin@gmail.com>
Add a function to tell runtime power management that the device is
starting in the off state instead of active or suspended.
Signed-off-by: Jordan Yates <jordan.yates@data61.csiro.au>
Add a Kconfig symbol to enable/disable power domain on Zephyr.
Disabling power domain save some memory / space.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
Add support for power domains on Zephyr. Power domains are implemented
as simple devices so they can use the existent Zephyr API, for resume
and suspend sync and async and also reference count.
The pm subsystem will ensure that domains are resumed before and
suspended after devices using them. For device runtime power
management, every time the device is got or released the same actions
is done to the domain it belongs.
As domains are implemented as simple devices, it is totally acceptable
a domain belongs to another domain.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
The usage count and device power state has to be restored when the
device action callback fails in the async operation. Otherwise it will
lead to an inconsistent state.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
The pm_device_runtime_enable did not suspend devices, so it assumed that
the device was in a physically suspended state. This change makes sure
that device is left in a suspended state if the device is initially
active.
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
In case runtime PM is not enabled (or not built-in), the get/put
functions always return 0 (instead of -ENOTSUP/-ENOSYS). When runtime PM
is disabled, a device is left into active state. Similarly, when device
runtime PM is not built-in, it is safe to assume that a device will
be active when it is called. If a user implements a custom solution, it
is its responsability to make sure that a device is active when using
it. For all these reasons, the -ENOTSUP/-ENOSYS are error codes that
should always be ignored by devices using get/put, since in practice it
means that: device is active, function is a no-op. The example below
illustrates how error handling is simplified:
```c
/* before: safe to ignore -ENOSYS/-ENOTSUP since device is active (we
* can continue)
*/
ret = pm_device_runtime_get(dev);
if ((ret < 0) && (ret != -ENOSYS) && (ret != -ENOTSUP)) {
return ret;
}
/* now */
ret = pm_device_runtime_get(dev);
if (ret < 0) {
return ret;
}
```
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
Change the function pm_device_runtime_enable() to return 0 on
success or an error code in case of error.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
Since enabling runtime pm on a device sets the device state to
suspended. The usage count has to be zeroed otherwise we may
have an inconsistency.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
It is not possible to rely on pm->dev to do lazy initialization of
some components. For example, in the follow sequence of commands:
1 - pm_device_runtime_enable()
pm->state == PM_DEVICE_STATE_SUSPENDED;
2 - pm_device_runtime_disable()
pm->state == PM_DEVICE_STATE_ACTIVE;
3 - pm_device_runtime_enable()
pm->state == PM_DEVICE_STATE_ACTIVE
After the first time pm_device_runtime_enable(), the device state will
be suspended, but after the second time this function executes the state
will be active. That is not consistent.
It is just easier to remove the branch check and always set those
fields for the sake of consistent.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
Although we are declaring `pm->enable`as bitfield, it ends up using
more memory due memory alignment.
Since we already have an atomic variable for device flags, this commit
adds a new flag to indicates whether or not device runtime is enabled.
Doing it we are saving some extra bits and avoiding need to lock the
mutex in several situations since we can atomically check if pm
runtime is enabled on a given device.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
Move all PM device runtime API calls from pm_device* to the
pm_device_runtime* namespace.
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
This patch refactors the runtime API to make it more clear and simple.
Relevant changes:
- API uses the action callback in a direct manner, avoiding unnecessary
overhead.
- API documentation has been improved to include detailed return error
codes.
- pm_runtime_disable() is now synchronous (to simplify possible error
paths) and returns error in case it fails. It is also safe to disable
in pre-kernel now.
- pm_runtime_put(_async)() will return -EALREADY if called with usage
count at zero (result of an unbalanced get/put call sequence)
- A transitional state has been added back. This makes code more
readable, and avoids using atomics (not required).
TODO:
- Solve in a better manner the asynchronous suspend error path (now
"solved" using asserts).
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
After the removal of pm_device_get_async, the pm_device_wait API has
become redundant. Its usage after pm_device_put_async should not be
considered a valid usecase, since after that call what will happen is a
pm_device_get (which is blocking).
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
As of today there is no clear usage of asynchronous gets, since in
general, a resume operation should be synchronous (we are about to use
the device immediately after resuming it). Removing this API simplifies
the runtime implementation in a significant way (refer to future
commits).
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
List of improvements:
- The PM logging module was only available if CONFIG_PM=y, however, it
was also used by Device PM (which can be selected without PM). A new
logging module has been created for Device PM.
- Log level is passed to LOG_MODULE_(DECLARE|REGISTER)
- Logger name has been adjusted to `pm` (was `power`)
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
- Rename to "action" to make its purpose more clear
- Use the _cb suffix to align with naming used for callbacks in other
areas.
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
Accessing members from pm_device improves code readability, since it
removes dev-> from most accesses.
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
Some includes were already performed by the device(_runtime).h header/s,
others like were not necessary.
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
PM_DEVICE_STATE_RESUMING and PM_DEVICE_STATE_SUSPENDING
are transitional states and are only used in device runtime. Remove it
and use device flag to keep track of a transition.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
The runtime API is referenced count and uses a uint32_t.
Avoid underflow when dealing with put requests.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
Return 0 in cases where the request does not trigger a device state
change, only incremented or to decremented the reference count.
Fixes#37821
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
The verb tense for the suspended state was not consistent with other
states. The likely reason: state was being used as a command/action.
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
If the device is already at the given state, do not call the device PM
control function. This makes sure that devices are only called to change
from one state to another.
Even though asynchronous device PM is completely broken, transitional
states are considered too.
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
The callback is not used anymore, so just delete it from the pm_control
callback signature.
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
The callback used by the device runtime PM can be easily replaced by a
simple state set after calling the state set/get calls. Broadcast logic
is simplified too, leading to the same previous behavior.
Since this is the only place where this callback was used, it can now be
removed from all devices and so pm_control callback signature
simplified.
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
Move all PM_DEVICE_STATE_* definitions to an enum. The
PM_DEVICE_STATE_SET and PM_DEVICE_STATE_GET definitions have been kept
out of the enum since they do not represent any state. However, their
name has not been changed since they will be removed soon.
All drivers and tests have been adjusted accordingly.
Signed-off-by: Gerard Marull-Paretas <gerard.marull@nordicsemi.no>
pm_device_runtime_state_set takes care of the check the reference count
and take the right action. It is not necessary check it in
pm_device_request.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
Assuming that pm_device_state_set is synchronous it is possible to
simplify the mutex usage. Now there are two places where the lock is
held, one in the worqueue handler and other in pm_device_request to
cover the synchronous path. It is no longer needed held the lock in the
pm_device_state_set callback and not needed to wait on the conditional
variable after set the state in the synchronous path.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
Most APIs have the default synchronous and an asynchronous version
with the sufix _async because that is the most common use.
All devices in tree right now are using the synchronous version, so
just change it to be consistent with the rest of the system.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
Since we are using mutex to protect critical sections and mutexes are
reentrant, it is possible to get rid of atomic for the state because
we can lock the mutex in device_pm_callback.
Signed-off-by: Flavio Ceolin <flavio.ceolin@intel.com>
